Road geometry extraction with fusion of low resolution satellite imagery and GPS trajectory using deep learning methods

Abstract

Road extraction is an important process which plays a crucial role in different applications such as improving navigation systems, facilitating urban planning and providing accurate road mapping in the deployment of autonomous vehicles, which are highly dependent on precise and reliable road information. This study examines the integration of Global Positioning System (GPS) trajectory data with low-resolution satellite imagery to enhance road detection techniques. The study focuses on the use of advanced convolutional neural network models, U-Net, ResUnet D-Linknet, which are tailored for semantic segmentation tasks in satellite images and novel fusion strategies to combine both satellite imagery and GPS trajectory data. Series of experiments conducted to evaluate the impact of different data fusion techniques and loss functions on the performance of the models. The study explores three main types of data fusion: early fusion, and three loss functions: Binary Cross-Entropy (BCE), Mean Squared Error (MSE) and Focal loss (FL). The results reveal that incorporating GPS data enhances road detection capabilities significantly, with late fusion providing the most substantial improvements. Among the tested models, ResUnet emerges as the most effective, particularly when employing a concatenation method for data fusion and utilizing MSE as the loss function. The study introduces the application of a new evaluation metric in the road detection domain, mBoundary-IoU (Mean Boundary Intersection Over Union), which provides a detailed assessment of extraction accuracy, particularly effective in accurately delineating the precise boundaries of road networks within urban landscapes. This metric is designed to complement the traditional Intersection over Union (IoU) by offering a more nuanced evaluation of the road outlines. A key part of the study involves the creation of a benchmark dataset that combines low-resolution satellite imagery with corresponding GPS data. This dataset covers Istanbul and Montreal. It is the first dataset of its kind in Turkey made available for public use and aims to facilitate comparative studies with satellite imagery and GPS trajectory and encourage further research into the integration of these two types of data. The research also investigates the variability in model performance across different geographic areas: Istanbul and Montreal. It is noted that the models exhibit better performance on the Montreal dataset, which features simpler and less congested road layouts compared to the complex and densely packed roads of Istanbul. This variability highlights the challenges and considerations needed when applying these models to different urban environments. In conclusion, the study demonstrates that the integration of GPS trajectory data with satellite imagery can significantly improve the precision and reliability of road detection systems. While the current findings are promising, the study suggests that further improvements could be achieved by exploring additional fusion techniques and by further customizing the deep learning models to accommodate the unique characteristics of different geographic areas.